Bootstrap-Ensembled Neural Networks for Robust Stellar Temperature and Age Predictions

This study presents a machine learning approach to predict the effective temperatures and their accuracy of premain sequence stars, which are essential for deriving stellar ages through isochrone fitting and starspot-dependent evolutionary models. We trained a Neural Network on high-quality spectroscopic temperatures from the Gaia-ESO Survey, using Gaia DR3 and 2MASS photometry as input features. To estimate predictive uncertainty, we implemented a Neural Network with bootstrap procedure, where each model was warm-started using the parameters of an initial network trained on the full dataset with K-fold cross-validation. This allowed accurate and robust temperature predictions for large stellar populations lacking spectroscopic data, with strong performance in the lowtemperature regime. The predicted temperatures were used to build Hertzsprung–Russell diagrams and derive stellar ages of young clusters by starspot evolutionary models, achieving good agreement with spectroscopic benchmarks and independent methods, such as lithium equivalent widths.